Decagon’s Playbook for Building Enterprise AI Applications

What It Covers

Decagon cofounders Jesse Zhang and Ashwin Srinivas explain how they shifted 90% of their AI workflow to open source models, why fine-tuned smaller models outperform frontier models on specific tasks, and how their enterprise AI agent evolved from customer support into a full business-process execution platform serving major banks, airlines, and telcos.

Key Questions Answered

• •Open Source Model Strategy: Decagon runs 90% of its workflow on fine-tuned open source models, reserving frontier models only for new or exploratory tasks. Fine-tuning smaller models on a single specific task — such as topic classification or bad-actor detection — produces better accuracy, lower latency, and lower cost than using a large general-purpose model. The false trade-off is intelligence versus cost; fine-tuned smaller models can outperform frontier models on targeted tasks across all three dimensions simultaneously.
• •Model Evaluation Framework: Evaluate AI models along three dimensions — cost, intelligence, and latency — and optimize for the limits your use case actually requires. Decagon's evals measure entire system performance end-to-end against customer outcomes, not individual model loss curves. Building proprietary benchmarks tied to specific production tasks is non-negotiable; public eval sets do not capture the nuance of real enterprise workflows and will produce misleading performance signals during fine-tuning.
• •Forward Deployed Engineering Discipline: Forward deployed engineers should produce core product improvements, not one-off customer customizations. Decagon's rule: every feature built during a customer engagement must benefit the next ten customers automatically. Companies that use forward deployed teams to execute bespoke work indefinitely become consulting firms, not scalable software companies. The forward deployed role is a temporary workflow-discovery mechanism that should collapse into product once the workflow is understood and repeatable.
• •Enterprise Sales Velocity: Decagon closes large enterprise contracts faster by mapping the deployment journey in granular detail before signing — including model risk governance, testing protocols, phased rollout sequencing, and issue-resolution processes. Enterprises buy when they can see a clear path from first meeting to 100% live deployment, not just a capable product. Jesse Zhang spends approximately 80% of his time on sales, focusing on shortening deployment timelines and building internal champions across large organizational hierarchies.
• •AI Concierge as Business Front Door: Decagon's long-term product thesis positions AI agents as the primary interface between a business and every customer interaction — reactive support, proactive outreach, and inbound sales qualification. The underlying capability enabling this expansion is improved instruction-following in newer models, which allows agents to handle open-ended, branching conversations rather than only tight, predefined paths. One customer went from deploying three agent journeys in a year with a competitor to seven journeys within one month after switching to Decagon.
• •Duet Autopilot — Agent Building Agents: Decagon's Duet Autopilot is a second, slower, frontier-model-powered agent that autonomously writes agent operating procedures, generates integration tools, creates test simulations, monitors live conversations, identifies underperforming topics, and drafts improvements — tasks previously requiring significant human engineering time. This architecture became viable only after reasoning models improved sufficiently. It compresses the time from new customer onboarding to a fully optimized live agent, and represents the productization of work that was originally done manually by forward deployed engineers.